Research On Copyright Algorithm Based On Adversarial Examples

Digital watermarking is one of the primary ways implementing copyright protection.The principle of digital watermarking is to embed specific,meaningful information on digital products,while in the meantime not affecting the regular usage of original carrier,and not being easily detected by users.Moreover,digital watermarks should be robust and can be traced under certain interferences.However,most of current digital watermarking techniques suffer from the conflict of imperceptibility and robustness.More specifically,larger amount of embedded information leads to higher robustness but lower imperceptibility,and vice versa.With the rapid development of deep learning,adversarial examples emerged.Adversarial examples interfere the output of deep learning models by adding a visually imperceptible perturbation to the original image.The generation of adversarial examples is very similar to digital watermarking techniques,and thus many related works on digital watermarking based on adversarial examples are proposed in recent years.This is a positive influence to the research on adversarial example,which makes the study of the combination of adversarial example and copyright protection a practical scenario.In this thesis,we first propose a novel adversarial example generation algorithm II-FGSM for one-stage object detector,and then propose a novel copyright protection system AdMarks in combination with II-FGSM and copyright protection algorithms.Finally,due to the lack of research on the frequency domain of adversarial examples,we improve II-FGSM on frequency domain so that AdMarks can achieve better performance.According to thorough experiments on real-life datasets,AdMarks significantly regulates the contradiction between imperceptibility and robustness of watermarks,and performs better than the state-of-the-art.The main contributions of this thesis are:1.Based on the one-stage object detection task,the targeted adversarial attack algorithm II-FGSM can quickly generate adversarial examples with better visual effects while maintaining a high attack success rate.2.AdMarks is a positive application of adversarial examples.More specifically,an adversarial attack algorithm II-FGSM is proposed to generate adversarial examples on a single-stage object detector,and copyright protection is then achieved through encoding and decoding detection information corresponding to the specific adversarial example.And a coding scheme e-NMI is designed for AdMarks to ensure uniqueness and recognizability of traceable watermarks.Sincedetection results are not publicly available and can only be detected,encoded,and traced by AdMarks,AdMarks thus shows a high level of security.3.Considering the lack of research of adversarial examples in frequency domain,we propose a classifier CNN-DCT based on DCT domain information,which improves performance of adversarial examples detection rate by 5%.In the meantime,the corresponding improved IAA-DCT algorithm resolves the drawbacks of adversarial examples in frequency domain.In addition,we further propose the II-FGSM-DCT algorithm,which improves the security of AdMarks significantly.